Contrastive Keyword Extraction (CKE) from Versioned Documents

Unsupervised Approach for Automatic Summarization (using Keywords) of Changes between two Document Versions

Introduction

Keyword Extraction (KE) in its original form is defined as the automatic identification of terms that best describe the subject of a document. KE has been successfully applied to many tasks as a special form of document summarization. Unlike traditional document summarization, in KE the created summary of a document does not consist of entire sentences but rather of a set of the most informative words or n-grams. In this work, we focus on a novel setting of keyword extraction that is applicable to versioned documents. In particular, we propose a novel task called Contrastive Keyword Extraction (CKE) which is defined as the summarization of changes between two versions of the same document. This is where Contrastive Keyword Extraction is fundamentally different from existing in-place diff tools, since it does not only extract differences between text versions, but it also ranks these differences based on how much they altered the meaning of the document. This could be especially useful when the amount of change is large (e.g., a substantial revision of a long document, such as a novel or a legal contract) and can be used to quickly summarize changes made to a document or even a collection of documents over time.

Try it yourself:

Open WebInterface »

Open Google Colab Notebook: »

Installing CKE

Requirements

Python3

Installation

To install CKE using pip:

pip install git+https://github.com/LukasEder1/ContrastiveKeywordExtraction

To upgrade using pip:

pip install git+https://github.com/LukasEder1/ContrastiveKeywordExtraction –-upgrade

Usage

Using default parameters

from cke import extract_contrastive_keywords

combined_kws, former_kws, latter_kws = extract_contrastive_keywords(document_a, document_b)

Parameters:

• document_a: Older Document Version (string)

• document_b: Newer Document Version (string)

Specifying more indepth parameters

from cke.sentence_comparision import match_sentences_semantic_search
from cke.sentence_importance import yake_weighted_importance
import string
import nltk

num_keywords = 10
max_ngram = 2
min_ngram = 1
threshold = 0.6
model = 'all-MiniLM-L6-v2'
num_splits = 1
symbols_to_remove = string.punctuation
stopwords = nltk.corpus.stopwords.words("english")

combined_kws, former_kws, latter_kws = extract_contrastive_keywords(document_a,
                                                                    document_b,
                                                                    num_keywords
                                                                    max_ngram=max_ngram,
                                                                    min_ngram=min_ngram,
                                                                    extra_stopwords=stopwords,
                                                                    importance_estimator= yake_weighted_importance,
                                                                    match_sentences=match_sentences_semantic_search,
                                                                    matching_model=model,
                                                                    threshold=threshold,
                                                                    symbols_to_remove=string.punctuation,
                                                                    num_splits=num_splits)

Parameters:

• document_a: Older Document Version (string)

• document_b: Newer Document Version (string)

• num_keywords: Number of Keywords to Extract (default=10)

• max_ngram: Maximum n-gram size of Keyphrases (default=2)

• min_ngram: Minimum n-gram size of Keyphrases (default=1)

• extra_stopwords: List of Stop words that should not be used as keywords (default=[])

• importance_estimator: Importance Calculator -> Predefined in cke.sentence_importance: yake_weighted_importance or text_rank_importance

• match_sentences: Sentence Matching Algorithm -> Predefined in cke.sentence_comparision: match_sentences_semantic_search or match_sentences_tfidf_weighted

• matching_model: Transformer Model for Semantic Search (default='all-MiniLM-L6-v2'): https://www.sbert.net/examples/applications/semantic-search/README.html

• threshold: Matching Threshold: acts as a lowerbound for whether or not two sentences should match (default=0.6)

• symbols_to_remove: List of Symbols that should be removed (defaul=[,])

• num_splits: Maximum Number Sentence a Source Sentence can split into (default=1)

Example

Using the first and the last revised version of the AP-News Article with id 17313 from the AP news-edits dataset (Google Drive)

Also available in the demo notebook (using documents[17313])

from cke import extract_contrastive_keywords

document_a, document_b = documents[17313]

combined_kws, former_kws, latter_kws = extract_contrastive_keywords(document_a, document_b, num_keywords=10, max_ngram=2)

Output

Combined Keywords

print(combined_kws)

{'state transportation': 0.16218195157227563,
 'transportation taxes': 0.16218195157227563,
 'new york': 0.11901325999183027,
 'records show': 0.11021500265104711,
 'york city': 0.09940959219076993,
 'city yellow': 0.09940959219076993,
 'taxes': 0.07831640607467383,
 'also sought': 0.06403374274208463,
 'attorneyclient privilege': 0.05739060614971311,
 'medallions': 0.047847894864559946}

Former Keywords

print(former_kws)

{'attorneyclient privilege': 0.14622844020560088,
 'fbi agents': 0.11396681642987444,
 'fire mueller': 0.107559144093367,
 'dead': 0.09032079989587966,
 'furious president': 0.09032079989587966,
 'president blasted': 0.09032079989587966,
 'blasted displeasure': 0.09032079989587966,
 'displeasure early': 0.09032079989587966,
 'early tuesday': 0.09032079989587966,
 'tuesday saying': 0.09032079989587966}

Latter Keywords

print(latter_kws)

{'state transportation': 0.1641493556133856,
 'transportation taxes': 0.1641493556133856,
 'new york': 0.12041330413087999,
 'records show': 0.11155200371377005,
 'york city': 0.10061551449904882,
 'city yellow': 0.10061551449904882,
 'taxes': 0.07926645022140402,
 'also sought': 0.0648105261203635,
 'medallions': 0.04842833023854907,
 'pleaded guilty': 0.0459996453501646}

Citation

If you use CKE in your own work please consider citing the following Paper. The Paper can be found at pdf.

Lukas Eder, Ricardo Campos and Adam Jatowt: Contrastive Keyword Extraction from Versioned Documents,
Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM 2023),
ACM Press, pp. 5026-5030 (2023)

@inproceedings{10.1145/3583780.3614735,
author = {Eder, Lukas and Campos, Ricardo and Jatowt, Adam},
title = {Contrastive Keyword Extraction from Versioned Documents},
year = {2023},
isbn = {9798400701245},
publisher = {Association for Computing Machinery},
url = {https://doi.org/10.1145/3583780.3614735},
doi = {10.1145/3583780.3614735},
booktitle = {Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},
pages = {5026–5030},
numpages = {5},
keywords = {keyword extraction, comparative summarization, change analysis},
location = {Birmingham, United Kingdom},
series = {CIKM '23}
}